Word Golf (2025)

Eric Xia & Julian Beaudry

word.golf is an online sport played with pretrained word embeddings. There are five rounds: in each, a source and target word with low cosine similarity are used. Word golf is played by finding the closest word out of a 4 by 7 grid: for instance, danger -> fragile -> falter -> linger is one path, which intuits ‘fragile’ is a bridge word between the source and target. When the target word enters the grid, it is made clearly visible to avoid frustrating the user. As a nod to the ‘golf’ aspect, there is always a two-click path from the source to the target.

Sam: When did you last launch, and what new features did you add?

Eric: We released major versions in June and July, adding persistent history, account functionality, and UI improvements. Julian: In July and early August, we added a catalog to replay past days and a race feature to compete with friends.

Sam: How did you distribute those updates?

Julian: At first we just told friends and played races with them. Existing users noticed the new “race” button at the top of the screen and started trying it.

Sam: How many active users do you have now?

Eric: About 30 daily active users, 200 weekly, and 500 over the past month.

Sam: Are most users strangers or friends?

Julian: Most daily players are users we don’t know personally; at least two-thirds. Some accounts are anonymous, so it’s hard to get exact numbers.

Sam: What does it cost to run Word Golf?

Julian: About one dollar a day for the app and database, both hosted on Digital Ocean. We’re far from capacity, so we can support many more users.

Sam: What are you working on next?

Julian: I’m building a prompt maker and revising the UI to make it less intimidating for new users. Eric: I’m experimenting with a new game using sentence embeddings, aiming to create enriching online experiences based on shared language knowledge.

Sam: How do you split the work?

Julian: Eric did a lot of the front-end. I handled authentication and back-end. I built the catalog; Eric built the race feature.

Sam: Were there any surprises or counterintuitive findings from the site analytics/data?

Eric: Directly reaching out to people with established user bases, with specific interests works better than passive advertising like posters.

Julian: We cut features like a skip button and a back button. They cluttered the UI, and we already had automatic skip after 12 jumps. We also had to completely rework the tutorial after testing with new players.

Sam: What’s been your favorite and least favorite part of the startup?

Eric: Favorite: seeing direct impact of my work on users. Least favorite: spending time on features that never ship.

Julian: My favorite moment was waking up to 100 new players overnight after the comic creator mentioned us. It’s rewarding to see people enjoy what we built.

Sam: What advice would you give other Brown students building products?

Eric: Be clear on why you’re building something and who it’s for. With Word Golf, I built it out of personal fascination with language vectors, so I was building for me.

Julian: Talk to users constantly. Quick feedback has been invaluable for us.

Sam: What could improve the summer program?

Eric: More interaction between different groups of fellows, to learn from each other’s struggles—especially on entrepreneurship.

Julian: Regular demo sessions, maybe monthly, so everyone can share progress and challenges. It would build a stronger community.